oz0 rollout start
For: Operators and developers looking up a command.
Start a rollout of a runtime version to a tenant's online nodes. Each step assigns the version to its nodes, waits until they run it, sends them their share of new jobs, and watches them for --soak. A step passes when its nodes stay healthy and jobs on the new version fail no more than on the old one; a step that fails rolls the upgraded nodes back by itself.
When fewer than five jobs end on the new version, the quality gate replays recent real inputs of the tenant's agents and flows there as shadow runs, whose tools that are not read-only answer from a stub (a harness agent's at the harness's permission bridge, a flow's steps with what they returned the first time), and compares how they and the few real jobs end with how they ended the first time.
Inputs that do not fit an agent's input schema are not replayed. Then the quality signals compare the new version's jobs of the step with the old one's: first-round outcome-check failures, cascade escalations, approvals people denied, sampled judges' failures and failures reported downstream. A signal more than its threshold worse rolls the step back; one with too few jobs on either side decides nothing; a signal never passes a step the other gates fail.
Usage
oz0 rollout start --runtime <version> [flags]
Flags
| Flag | Type | Default | Description |
|---|---|---|---|
--batch | int32 | half of the rest | nodes in each later step |
--canary | int32 | 1 | nodes in the first step |
--json | bool | print JSON instead of a table | |
--quality-cascade | points | how many points more of the candidate's jobs may go to a later model of their cascade before the step fails and the candidate rolls back (0 to 1, 0.1 is 10 points, or off; default the plugin's configuration, else 0.1; only stricter) | |
--quality-denials | points | how many points more of the candidate's approvals people may deny before the step fails and the candidate rolls back (0 to 1, 0.1 is 10 points, or off; default the plugin's configuration, else 0.1; only stricter) | |
--quality-min-jobs | int32 | the plugin's configuration, else 10; only lower | the fewest jobs each version needs in a quality signal for it to count, 1 to 1000 |
--quality-outcome | points | how many points more of the candidate's jobs than the installed version's may fail their outcome check's first round before the step fails and the candidate rolls back (0 to 1, 0.1 is 10 points, or off; default the plugin's configuration, else 0.1; only stricter) | |
--quality-reported | points | how many points more of the candidate's reported results may be failures before the step fails and the candidate rolls back (0 to 1, 0.1 is 10 points, or off; default the plugin's configuration, else 0.1; only stricter) | |
--quality-sampled | points | how many points more of the candidate's sampled jobs a sampled judge may fail before the step fails and the candidate rolls back (0 to 1, 0.1 is 10 points, or off; default the plugin's configuration, else 0.1; only stricter) | |
--runtime | string | the runtime version to move to: one the edge serves | |
--shadow-runs | int32 | 5 | recent real inputs the quality gate replays on the new version, side effects stubbed, when too few real jobs end there; 0 for none, at most 20 |
--soak | duration | 1m | how long each step is watched before its gates |
--start-timeout | duration | 5m | how long a step's nodes may take to run the new version |
--tenant | string | default | tenant |
Inherited flags
| Flag | Type | Default | Description |
|---|---|---|---|
--context | string | $OZ0_CONTEXT or <config dir>/orchestrator-zero/context.json | CLI context file written by server init |
--log-level | string | info | log level: debug, info, warn or error |
See all oz0 commands and the global flags.